search

JingyunLiang / MANet

Official PyTorch code for Mutual Affine Network for Spatially Variant Kernel Estimation in Blind Image Super-Resolution (MANet, ICCV2021)
https://arxiv.org/abs/2108.05302
Apache License 2.0
169 stars 23 forks source link

[Can estimated kernels(spatially variant) be used to deconvolution debluring?] #4

Closed CaptainEven closed 3 years ago

CaptainEven commented 3 years ago

Can estimated kernels(spatially variant) be applied to deconvolution debluring explicitly, using FFT and wiener filter?

JingyunLiang commented 3 years ago

It is possible, but I didn't try it. Note that for deconvolution, the kernels are often much more complex (e.g., motion blur). It may be harder to estimate the kernels accurately, though one good thing is that there is no downscaling in deconvolution. We can summarize them as follows:

image super-resolution: downscaling (hard, much loss of information) + blur (often assumed to be simple, at most anisotropic Gaussian) + noise

image deblurring: blur only (often motion blur kernels, as Gaussian blurs are too easy for this task, even traditional methods may solve it well) + noise

CaptainEven commented 3 years ago

@JingyunLiang Thanks for the explanation! I got it, and i 'll try it maybe when i'm free .